[This is Kevin's high-confidence prediction, made on the 12/22/23 episode of Hard Fork]
This question asks whether a lawless LLM chatbot will get to 10 million daily active users by the end of 2024. For the purposes of resolving this market, "lawless" should be construed to mean "significantly less restrained than comparable models from OpenAI, Google and Anthropic." It does not have to have literally zero rules to qualify as lawless, but a lawless chatbot should be able to respond to most queries and prompts (even explicit/violent/sexual ones) without refusals.
Related questions
Love this and the podcast generally. I bet that over the next year similar actions will occur with an LLM. In fact this is already close to happening, so a year is a very generous time.
Some things im thinking about:
A very strong model has already been leaked, the LLAMA model from META. As more and more companies want to have a competing model the odds of a leak of one (or more) of these prebaked models increases. All it takes is an enterprising person to slap it together into a model and it'll work.
While Bard and ChatGPT have strong filtering, more and more companies are using these models that do not have the same budgets. Consider the SnapChat AI which a few months ago which users figured out could see their location. A bit of an obvious oversight that probably would have come out sooner had Snap not been trying to ship anything with AI since they haven't been relevant in years. (Sorry.) As the market for these "second tier" LLMs increase, so do the possibility of abuse, and thus, a payout on this.
An older method (but still checks out) I use to see some cutting edge (ab?)use of these tools is 4chan. Over the past year the site has become a haven for use of AI generated image content, mainly pornography, as users push the boundaries of that tech. This to me signals there is a strong market, and base of users, willing and able to push the boundaries of LLMs.
You know what 4chan is go look it up urself.
We are already seeing sites like RizzGPT play with the idea users can talk with characters. And "lawless" in my mind can include an LLM that will engage with the user's sexual prompts.
10 million isn’t that much counting International, who will be especially allergic to American centric, norms embodied in the leading models.
Clarifying question: what is it refuses to talk about politically sensitive topics in its host nation, but is otherwise uncensored according to the list of topics in the current description?
@0482 Possible, sure, but could lead to verification issues if the counting method and source data is deemed “proprietary”